ledger-signals: extend duplicateDirections to see still-open PRs

[ruvnet]

Nightly Dream Cycle, 2026-08-26. DEEP=ledger-signals, SCAN=witness,verify. Full report: docs/dream-cycle/2026-08-26-ledger-signals-report.md. Issue: #32.

Hypothesis

Given duplicateDirections only scans ledger rows already merged into LEDGER.md on main, when learningSignals() gains an optional pendingFindings: string[] parameter folded into the same first-6-word normalization/counting pipeline already used for merged rows, then a direction repeated across merged rows + still-open PRs' findings (>= 3 combined) should be flagged — while the default (omitted) path stays byte-identical. Frozen before implementation.

Candidate

+130/-9 across 5 files (1 new report, 4 edits: packages/ledger/src/index.ts + its tests, packages/cli/src/index.ts + its tests). One conceptual change: an additive pendingFindings option on learningSignals() + a --pending "f1|f2" CLI flag on ledger signals. No removal, no behavior change when the option/flag is omitted.

Evaluation Receipt

Real evaluator: npm test (vitest, this repo's own bench entrypoint).

	Baseline (parent, commit 074fd1f)	Candidate
Tests	96	104 (+8, 0 removed, 0 modified)
Result	96 passed	104 passed

Baseline reconfirmed via git stash immediately before finalizing (96/96, exact match). Live receipt:

$ node packages/cli/dist/bin.js ledger signals --path /tmp/synthetic-ledger.md
{ "duplicateDirections": [], ... }   # 1 merged row only, below threshold

$ node packages/cli/dist/bin.js ledger signals --path /tmp/synthetic-ledger.md \
    --pending "zero merge streak signal ignores real pr merge state (cli --merged flag night)|zero merge streak signal ignores real pr merge state (tonight's near-duplicate before catching #15/#27)"
{ "duplicateDirections": ["zero merge streak signal ignores real"], ... }

Baseline

Parent commit 074fd1fe6ceac4bebc993be6b6b0f0f69cc750f5, 96/96 tests passing, clean build (npm ci && npm run build, no wasm/NAPI packages in this workspace).

Darwin Lineage

Not run — DARWIN=not-applicable. Small, pure, deterministic string-processing change; no evolvable population. Same judgment as the 2026-08-13/2026-08-24 precedents.

Evidence

Confirmed live via GitHub MCP tonight: LEDGER.md on main has 1 row while GitHub has 9 open, unmerged dream-cycle draft PRs (#15, #17, #19, #21, #23, #24, #27, #29, #30), 0 merged since #7/#13. PR #15 (2026-08-16) and PR #27 (2026-08-24) independently propose closely related fixes for the same zeroMergeStreak-can't-see-real-merges defect — a real, demonstrated duplicate-direction case the existing detector cannot see, because neither finding ever landed in LEDGER.md.

Honesty correction (post-critique): the draft report initially implied this candidate "would have caught #15 vs #27." Verified directly (directionKey on the real title text) that it would not — their real titles don't share a 6-word lexical prefix. The candidate extends recall (pending findings become visible to the same detector), not the detector's underlying matching precision, which is a known, pre-existing limitation unchanged by this PR. See the full report for the corrected, honest framing.

Reward-Hack Check

Independent critic: a fresh subagent (Agent tool, no access to this candidate's authoring context) reviewed the diff adversarially. Initial verdict: BLOCKED on two findings:

1. parsePendingFindings split on , — the real motivating PR developer-experience: thread real merge state into zero-merge learning signal (cli, tui) #27 title contains a comma and would fracture into a bogus extra finding. Not covered by the original tests. Fixed: delimiter switched to | (matches LEDGER.md's own field separator), regression test added.
2. The draft report overstated that this candidate "would have caught ledger-signals: derive zeroMergeStreak from Prior-night fates tokens #15 vs developer-experience: thread real merge state into zero-merge learning signal (cli, tui) #27." Fixed: report corrected to the honest, narrower claim (see Evidence above).

Both fixes verified: rebuilt, reran npm test (104/104, 0 regressions), re-ran parsePendingFindings against the real PR #27 title to confirm the comma is preserved. Re-reviewed after fixes: CLEAR. No pre-existing test weakened/removed; default path verified byte-identical via a dedicated regression test; no new I/O/network/shell-exec/credential surface; no threshold/gate/safety constant touched; scope limited to the 4 intended source+test files.

Security Review

Pure string parsing (parsePendingFindings, directionKey, bumpDirection) — no eval, no shell interpolation, no external I/O, no LLM calls. --pending's input is always caller-supplied, never attacker-controllable runtime input.

Regression Analysis

0 pre-existing tests modified or removed. All 96 baseline tests still pass unchanged; 8 new tests added (3 in packages/ledger, 5 in packages/cli, including the post-critique comma-preservation regression test).

ADR

None — additive, opt-in extension to an existing, already-tested signal-computation function, not a new architectural decision. Same judgment as the PR #15/#27 precedents for adjacent changes.

Gist

No gh gist create / gh CLI available in this environment. Report committed at docs/dream-cycle/2026-08-26-ledger-signals-report.md instead. GIST=LOCAL.

Issue

#32

Witness

report_sha256 : c75b4a90f6ddab5ab26d0d547e36be1455916f748c559681c48fe130367f43fc
session_commit: 074fd1fe6ceac4bebc993be6b6b0f0f69cc750f5
witness       : 7bf7a1d1c628430d4a91860a8df4ca0e569b1f4c4007e49e790f996cf23619dc

Verify: sha256sum docs/dream-cycle/2026-08-26-ledger-signals-report.md, then printf '%s%s' "<that hash>" "074fd1fe6ceac4bebc993be6b6b0f0f69cc750f5" | sha256sum must equal the witness above. Confirmed tonight via dream-machine witness verify (✓ VALID).

Merge Policy

Draft — human review required. This PR does not carry the automerge-safe label: it modifies a decision-support signal (duplicateDirections) that future nights' STEP 1.1 relies on, which is exactly the low-confidence-blast-radius case this repo's guarded-auto-merge policy (.github/workflows/automerge.yml) keeps human-review-only. The session never applies that label itself and never merges.

Also flagged for human triage, not fixed by this PR: 9 dream-cycle draft PRs are open and unmerged (#15, #17, #19, #21, #23, #24, #27, #29, #30). This candidate makes the symptom (undetected duplicate directions) more visible for a future night that wires --pending in, but the actual fix — merging or closing the backlog — is a human decision outside any single nightly candidate's scope.

---

Generated by Claude Code

[ruvnet]

Resolved a merge conflict against main (which advanced with PR #24 + several "portfolio 30x" ledger rows from a separate cross-repo process) by keeping all of main's rows and appending this branch's 2026-08-26 row after them — no row content changed, append-only. Rebuilt, relinted, and reran the full suite post-merge: 106/106 passing, 0 regressions.

Side observation while re-verifying: dream-machine ledger verify reports 9 schema errors, all in the pre-existing "portfolio 308–311" rows (2026-08-22 through 2026-08-26), which use compound verdicts (e.g. ACCEPT / INCONCLUSIVE) and evaluated: partial — neither valid per this repo's own Verdict/Evaluated enums (packages/ledger/src/index.ts). Confirmed this already fails on origin/main by itself, before this merge — not introduced by this PR, and this branch's own new row is schema-valid. Left untouched; out of this PR's scope to rewrite another process's committed ledger rows, flagging for maintainer awareness.

---

Generated by Claude Code